Merge tag 'misc-habanalabs-next-2022-11-23' of https://git.kernel.org/pub/scm/linux/kernel/git/ogabbay/linux into char-misc-next

Oded writes:

This tag contains habanalabs driver changes for v6.2:

- New feature of graceful hard-reset. Instead of immediately killing the
  user-process when a command submission times out, we wait a bit and give
  the user-process notification and let it try to close things gracefully,
  with the ability to retrieve debug information.

- Enhance the EventFD mechanism. Add new events such as access to illegal
  address (RAZWI), page fault, device unavailable. In addition, change the
  event workqueue to be handled in a single-threaded workqueue.

- Allow the control device to work during reset of the ASIC, to enable
  monitoring applications to continue getting the data.

- Add handling for Gaudi2 with PCI revision 2.

- Reduce severity of prints due to power/thermal events.

- Change how we use the h/w to perform memory scrubbing in Gaudi2.

- Multiple bug fixes, refactors and renames.

* tag 'misc-habanalabs-next-2022-11-23' of https://git.kernel.org/pub/scm/linux/kernel/git/ogabbay/linux: (63 commits)
  habanalabs: fix VA range calculation
  habanalabs: fail driver load if EEPROM errors detected
  habanalabs: make print of engines idle mask more readable
  habanalabs: clear non-released encapsulated signals
  habanalabs: don't put context in hl_encaps_handle_do_release_sob()
  habanalabs: print context refcount value if hard reset fails
  habanalabs: add RMWREG32_SHIFTED to set a val within a mask
  habanalabs: fix rc when new CPUCP opcodes are not supported
  habanalabs/gaudi2: added memset for the cq_size register
  habanalabs: added return value check for hl_fw_dynamic_send_clear_cmd()
  habanalabs: increase the size of busy engines mask
  habanalabs/gaudi2: change memory scrub mechanism
  habanalabs: extend process wait timeout in device fine
  habanalabs: check schedule_hard_reset correctly
  habanalabs: reset device if still in use when released
  habanalabs/gaudi2: return to reset upon SM SEI BRESP error
  habanalabs/gaudi2: don't enable entries in the MSIX_GW table
  habanalabs/gaudi2: remove redundant firmware version check
  habanalabs/gaudi: fix print for firmware-alive event
  habanalabs: fix print for out-of-sync and pkt-failure events
  ...

Author: gregkh

Documentation/ABI/testing/debugfs-driver-habanalabs

@@ -91,6 +91,13 @@ Description:    Enables the root user to set the device to specific state.
                 Valid values are "disable", "enable", "suspend", "resume".
                 User can read this property to see the valid values
 
+What:           /sys/kernel/debug/habanalabs/hl<n>/device_release_watchdog_timeout
+Date:           Oct 2022
+KernelVersion:  6.2
+Contact:        [email protected]
+Description:    The watchdog timeout value in seconds for a device relese upon
+                certain error cases, after which the device is reset.
+
 What:           /sys/kernel/debug/habanalabs/hl<n>/dma_size
 Date:           Apr 2021
 KernelVersion:  5.13

drivers/misc/habanalabs/common/command_submission.c

@@ -742,13 +742,11 @@ static void cs_do_release(struct kref *ref)
 		 */
 		if (hl_cs_cmpl->encaps_signals)
 			kref_put(&hl_cs_cmpl->encaps_sig_hdl->refcount,
-						hl_encaps_handle_do_release);
+					hl_encaps_release_handle_and_put_ctx);
 	}
 
-	if ((cs->type == CS_TYPE_WAIT || cs->type == CS_TYPE_COLLECTIVE_WAIT)
-			&& cs->encaps_signals)
-		kref_put(&cs->encaps_sig_hdl->refcount,
-					hl_encaps_handle_do_release);
+	if ((cs->type == CS_TYPE_WAIT || cs->type == CS_TYPE_COLLECTIVE_WAIT) && cs->encaps_signals)
+		kref_put(&cs->encaps_sig_hdl->refcount, hl_encaps_release_handle_and_put_ctx);
 
 out:
 	/* Must be called before hl_ctx_put because inside we use ctx to get
@@ -798,7 +796,7 @@ static void cs_do_release(struct kref *ref)
 static void cs_timedout(struct work_struct *work)
 {
 	struct hl_device *hdev;
-	u64 event_mask;
+	u64 event_mask = 0x0;
 	int rc;
 	struct hl_cs *cs = container_of(work, struct hl_cs,
 						 work_tdr.work);
@@ -830,11 +828,7 @@ static void cs_timedout(struct work_struct *work)
 	if (rc) {
 		hdev->captured_err_info.cs_timeout.timestamp = ktime_get();
 		hdev->captured_err_info.cs_timeout.seq = cs->sequence;
-
-		event_mask = device_reset ? (HL_NOTIFIER_EVENT_CS_TIMEOUT |
-				HL_NOTIFIER_EVENT_DEVICE_RESET) : HL_NOTIFIER_EVENT_CS_TIMEOUT;
-
-		hl_notifier_event_send_all(hdev, event_mask);
+		event_mask |= HL_NOTIFIER_EVENT_CS_TIMEOUT;
 	}
 
 	switch (cs->type) {
@@ -869,8 +863,12 @@ static void cs_timedout(struct work_struct *work)
 
 	cs_put(cs);
 
-	if (device_reset)
-		hl_device_reset(hdev, HL_DRV_RESET_TDR);
+	if (device_reset) {
+		event_mask |= HL_NOTIFIER_EVENT_DEVICE_RESET;
+		hl_device_cond_reset(hdev, HL_DRV_RESET_TDR, event_mask);
+	} else if (event_mask) {
+		hl_notifier_event_send_all(hdev, event_mask);
+	}
 }
 
 static int allocate_cs(struct hl_device *hdev, struct hl_ctx *ctx,
@@ -1011,6 +1009,34 @@ static void cs_rollback(struct hl_device *hdev, struct hl_cs *cs)
 		hl_complete_job(hdev, job);
 }
 
+/*
+ * release_reserved_encaps_signals() - release reserved encapsulated signals.
+ * @hdev: pointer to habanalabs device structure
+ *
+ * Release reserved encapsulated signals which weren't un-reserved, or for which a CS with
+ * encapsulated signals wasn't submitted and thus weren't released as part of CS roll-back.
+ * For these signals need also to put the refcount of the H/W SOB which was taken at the
+ * reservation.
+ */
+static void release_reserved_encaps_signals(struct hl_device *hdev)
+{
+	struct hl_ctx *ctx = hl_get_compute_ctx(hdev);
+	struct hl_cs_encaps_sig_handle *handle;
+	struct hl_encaps_signals_mgr *mgr;
+	u32 id;
+
+	if (!ctx)
+		return;
+
+	mgr = &ctx->sig_mgr;
+
+	idr_for_each_entry(&mgr->handles, handle, id)
+		if (handle->cs_seq == ULLONG_MAX)
+			kref_put(&handle->refcount, hl_encaps_release_handle_and_put_sob_ctx);
+
+	hl_ctx_put(ctx);
+}
+
 void hl_cs_rollback_all(struct hl_device *hdev, bool skip_wq_flush)
 {
 	int i;
@@ -1039,6 +1065,8 @@ void hl_cs_rollback_all(struct hl_device *hdev, bool skip_wq_flush)
 	}
 
 	force_complete_multi_cs(hdev);
+
+	release_reserved_encaps_signals(hdev);
 }
 
 static void
@@ -2001,6 +2029,8 @@ static int cs_ioctl_reserve_signals(struct hl_fpriv *hpriv,
 	 */
 	handle->pre_sob_val = prop->next_sob_val - handle->count;
 
+	handle->cs_seq = ULLONG_MAX;
+
 	*signals_count = prop->next_sob_val;
 	hdev->asic_funcs->hw_queues_unlock(hdev);
 
@@ -2350,10 +2380,8 @@ static int cs_ioctl_signal_wait(struct hl_fpriv *hpriv, enum hl_cs_type cs_type,
 	/* We finished with the CS in this function, so put the ref */
 	cs_put(cs);
 free_cs_chunk_array:
-	if (!wait_cs_submitted && cs_encaps_signals && handle_found &&
-							is_wait_cs)
-		kref_put(&encaps_sig_hdl->refcount,
-				hl_encaps_handle_do_release);
+	if (!wait_cs_submitted && cs_encaps_signals && handle_found && is_wait_cs)
+		kref_put(&encaps_sig_hdl->refcount, hl_encaps_release_handle_and_put_ctx);
 	kfree(cs_chunk_array);
 out:
 	return rc;

drivers/misc/habanalabs/common/context.c

@@ -9,38 +9,46 @@
 
 #include <linux/slab.h>
 
-void hl_encaps_handle_do_release(struct kref *ref)
+static void encaps_handle_do_release(struct hl_cs_encaps_sig_handle *handle, bool put_hw_sob,
+					bool put_ctx)
 {
-	struct hl_cs_encaps_sig_handle *handle =
-		container_of(ref, struct hl_cs_encaps_sig_handle, refcount);
 	struct hl_encaps_signals_mgr *mgr = &handle->ctx->sig_mgr;
 
+	if (put_hw_sob)
+		hw_sob_put(handle->hw_sob);
+
 	spin_lock(&mgr->lock);
 	idr_remove(&mgr->handles, handle->id);
 	spin_unlock(&mgr->lock);
 
-	hl_ctx_put(handle->ctx);
+	if (put_ctx)
+		hl_ctx_put(handle->ctx);
+
 	kfree(handle);
 }
 
-static void hl_encaps_handle_do_release_sob(struct kref *ref)
+void hl_encaps_release_handle_and_put_ctx(struct kref *ref)
 {
 	struct hl_cs_encaps_sig_handle *handle =
-		container_of(ref, struct hl_cs_encaps_sig_handle, refcount);
-	struct hl_encaps_signals_mgr *mgr = &handle->ctx->sig_mgr;
+			container_of(ref, struct hl_cs_encaps_sig_handle, refcount);
 
-	/* if we're here, then there was a signals reservation but cs with
-	 * encaps signals wasn't submitted, so need to put refcount
-	 * to hw_sob taken at the reservation.
-	 */
-	hw_sob_put(handle->hw_sob);
+	encaps_handle_do_release(handle, false, true);
+}
 
-	spin_lock(&mgr->lock);
-	idr_remove(&mgr->handles, handle->id);
-	spin_unlock(&mgr->lock);
+static void hl_encaps_release_handle_and_put_sob(struct kref *ref)
+{
+	struct hl_cs_encaps_sig_handle *handle =
+			container_of(ref, struct hl_cs_encaps_sig_handle, refcount);
 
-	hl_ctx_put(handle->ctx);
-	kfree(handle);
+	encaps_handle_do_release(handle, true, false);
+}
+
+void hl_encaps_release_handle_and_put_sob_ctx(struct kref *ref)
+{
+	struct hl_cs_encaps_sig_handle *handle =
+			container_of(ref, struct hl_cs_encaps_sig_handle, refcount);
+
+	encaps_handle_do_release(handle, true, true);
 }
 
 static void hl_encaps_sig_mgr_init(struct hl_encaps_signals_mgr *mgr)
@@ -49,20 +57,22 @@ static void hl_encaps_sig_mgr_init(struct hl_encaps_signals_mgr *mgr)
 	idr_init(&mgr->handles);
 }
 
-static void hl_encaps_sig_mgr_fini(struct hl_device *hdev,
-			struct hl_encaps_signals_mgr *mgr)
+static void hl_encaps_sig_mgr_fini(struct hl_device *hdev, struct hl_encaps_signals_mgr *mgr)
 {
 	struct hl_cs_encaps_sig_handle *handle;
 	struct idr *idp;
 	u32 id;
 
 	idp = &mgr->handles;
 
+	/* The IDR is expected to be empty at this stage, because any left signal should have been
+	 * released as part of CS roll-back.
+	 */
 	if (!idr_is_empty(idp)) {
-		dev_warn(hdev->dev, "device released while some encaps signals handles are still allocated\n");
+		dev_warn(hdev->dev,
+			"device released while some encaps signals handles are still allocated\n");
 		idr_for_each_entry(idp, handle, id)
-			kref_put(&handle->refcount,
-					hl_encaps_handle_do_release_sob);
+			kref_put(&handle->refcount, hl_encaps_release_handle_and_put_sob);
 	}
 
 	idr_destroy(&mgr->handles);

drivers/misc/habanalabs/common/debugfs.c

@@ -1769,6 +1769,11 @@ void hl_debugfs_add_device(struct hl_device *hdev)
 				dev_entry,
 				&hl_timeout_locked_fops);
 
+	debugfs_create_u32("device_release_watchdog_timeout",
+				0644,
+				dev_entry->root,
+				&hdev->device_release_watchdog_timeout_sec);
+
 	for (i = 0, entry = dev_entry->entry_arr ; i < count ; i++, entry++) {
 		debugfs_create_file(hl_debugfs_list[i].name,
 					0444,